Novel electric power system weak link identification method based on deep learning

Abstract

The invention discloses a novel electric power system weak link identification method based on deep learning, and relates to the technical field of electric power system analysis, and the method comprises the steps: identifying a node energy transmission abnormal mode through the sudden change of phase interaction coupling intensity; training a deep learning model according to a small number of historical disturbance samples; inputting the node energy transmission abnormal mode into a recognition deep learning model to obtain a node response path activation sequence; constructing a system state manifold space based on the node response path activation sequence; when local distortion occurs in the manifold space, the distortion feature feedback is used for adjusting the phase interaction coupling intensity mutation recognition; generating a node vulnerability topology fingerprint according to the adjusted geometric topology characteristics of the coupling strength mutation and manifold distortion; recognizing weak links of the system through topological fingerprints; according to the method, through fusion of phase interactive coupling analysis, a meta-learning algorithm and a manifold feedback mechanism, adaptive weak link accurate identification under the condition of few samples is realized.

Description

Technical Field

[0001] This invention relates to the field of power system analysis technology, specifically to a novel method for identifying weak links in power systems based on deep learning. Background Technology

[0002] Identifying weak links in the power system is a key technology for ensuring the safe and stable operation of the power grid. Weak links are prone to failure or instability when disturbed, and accurate identification is of great significance for preventing cascading failures.

[0003] Existing technologies mainly include static methods based on power flow analysis, dynamic methods based on voltage stability, and intelligent methods based on machine learning. Power flow analysis-based methods identify faults by calculating static parameters such as node voltage and branch power. While computationally simple, these methods fail to reflect the dynamic characteristics of the system and have limited effectiveness in identifying rapid disturbances. Voltage stability-based methods identify faults by analyzing voltage stability margins, reflecting quasi-steady-state dynamic characteristics, but their ability to handle sudden disturbances is insufficient. Machine learning-based methods perform pattern recognition by training models, possessing nonlinear modeling capabilities, but require a large number of training samples. However, power system fault samples are scarce and costly to obtain, limiting the model's generalization ability.

[0004] The existing technology has the following shortcomings: static analysis methods cannot capture transient response features during disturbances; the sample requirements of traditional deep learning methods do not match the scarcity of fault samples; existing methods lack adaptive adjustment mechanisms and are difficult to adapt to changes in operating conditions; feature representation is singular and fails to integrate local anomalies and global state features; physical mechanisms are not fully utilized, and the reliability of identification needs to be improved.

[0005] Therefore, there is a need for an identification method that can accurately identify weak links under conditions of few samples, has adaptive capabilities, and makes full use of the physical characteristics of the power system. Summary of the Invention

[0006] The purpose of this invention is to provide a novel power system weak link identification system and method based on deep learning to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a novel deep learning-based method for identifying weak links in power systems, comprising: S101: Collects the active and reactive power of each node in the power system; identifies abnormal energy transmission patterns of nodes by abrupt changes in the phase interaction coupling strength between active and reactive power. S102: Based on a small number of historical disturbance samples, a deep learning model for identifying weak links is generated by training a meta-learning algorithm; the abnormal energy transmission pattern of nodes is input into the deep learning model for identifying weak links to obtain the node response path activation sequence. S103: Construct the system state manifold space based on the temporal evolution characteristics of the node response path activation sequence; when the manifold space is locally distorted, use the distortion characteristics to feedback and adjust the abrupt identification of the phase interaction coupling strength in step S101. S104: Generate node vulnerability topological fingerprints based on the geometric topological features of the adjusted coupling strength abrupt change and manifold distortion; identify node clusters with the same vulnerable topological structure as weak links in the system by isomorphic matching of the topological fingerprints.

[0008] Furthermore, the method of identifying abnormal node energy transfer modes through abrupt changes in the phase interaction coupling strength between active and reactive power includes: Construct a complex power vector from active power and reactive power, and establish a power complex vector time series. A phase jump detection algorithm is used to identify phase jump events in the power complex vector time series, and a set of phase jump events is obtained. Calculate the magnitude of the coupling strength change for each abrupt event in the set of phase abrupt events to form a coupling strength change spectrum; Extract the time window and amplitude features corresponding to abnormal peaks from the coupling strength variation spectrum; Anomaly patterns in node energy transmission are constructed based on the time window and amplitude characteristics of abnormal peak values.

[0009] Furthermore, the step of identifying phase abrupt change events in the power complex vector time series using the phase difference jump detection algorithm includes: Calculate the phase angle difference between adjacent time steps of the power complex vector time series, and generate a phase angle difference sequence; Establish a statistical distribution model for the phase angle difference sequence and determine the discrimination threshold for phase jump; The time corresponding to the phase angle difference exceeding the discrimination threshold is marked as a candidate point for phase change; Verify that the phase continuity before and after the phase change candidate point is indeed interrupted, confirm it as a real phase change event, and output the set of phase change events.

[0010] Furthermore, the step of training a deep learning model for weak link identification using a meta-learning algorithm based on a small number of historical disturbance samples includes: A small number of historical perturbation samples are divided into a support set and a query set to construct a meta-learning training task. Design a two-layer optimized meta-learning network, which includes a fast-adaptive inner deep network and a cross-task generalizing outer deep network; Train the parameters of the inner deep network on the support set and optimize the meta-parameters of the outer deep network on the query set; Through multiple rounds of meta-training iterations, the two-layer deep network is equipped with the ability to quickly learn and identify weak points from a small number of new samples; Save the parameters and structure configuration of the two-layer deep network after training convergence to generate a deep learning model for weak link identification.

[0011] Furthermore, the step of inputting the abnormal node energy transmission pattern into the weak link identification deep learning model to obtain the node response path activation sequence includes: Convert the node energy transfer anomaly pattern into the input tensor format of the weak link identification deep learning model; Extract multi-level feature representations of abnormal patterns through convolutional and pooling layers of deep learning models; Multi-level feature representations are input into the recurrent neural network layers of the model to predict possible response propagation paths; Calculate the activation probability and activation intensity of nodes on each predicted path; The predicted path and activation information are combined to form the node response path activation sequence.

[0012] Furthermore, the extraction of multi-level feature representations of abnormal patterns through convolutional and pooling layers of the deep learning model includes: The input tensor is passed through the first convolutional layer to perform feature convolution operations, generating a primary feature map. Perform max pooling on the primary feature map to reduce the feature dimension while preserving salient features; The pooling results are then passed through a second convolutional layer for deep feature extraction to generate a high-level feature map. The high-level feature map is flattened into a one-dimensional feature vector, and the output is a multi-level feature representation of the anomaly pattern.

[0013] Furthermore, the construction of the system state manifold space based on the temporal evolution characteristics of the node response path activation sequence includes: Collect node response path activation sequences at multiple consecutive time points to form an activation sequence time-domain dataset; The activation sequence time-domain dataset is mapped to a low-dimensional manifold space using a manifold embedding algorithm to obtain a low-dimensional manifold space coordinate representation. Identify the distribution patterns and clustering structures of data points in a low-dimensional manifold space; Extract the geometric boundaries and internal connectivity of the cluster structure to construct the system state manifold space.

[0014] Furthermore, the step of mapping the activation sequence time-domain dataset to a low-dimensional manifold space using a manifold embedding algorithm includes: Construct a proximity graph among data points in the time-domain dataset of the activation sequence; Calculate the Laplacian matrix of the proximity graph, perform eigenvalue decomposition, and obtain the eigenvalue vector and eigenvector matrix; Select the eigenvectors corresponding to the smallest non-zero eigenvalues ​​as the embedding basis vectors; The original data points are projected onto the subspace spanned by the embedded basis vectors to obtain a low-dimensional manifold space coordinate representation.

[0015] Further, the step of generating node vulnerability topological fingerprints based on the adjusted coupling strength abrupt change and manifold distortion geometric topological features includes: Extract the amplitude and direction vectors of the abrupt change in the phase interaction coupling strength after adjustment; Obtain the Gaussian curvature and mean curvature values ​​of the manifold distortion region; Calculate the tensor product of the magnitude vector and the Gaussian curvature to obtain the first topological feature tensor; The second topological feature vector is obtained by calculating the outer product of the direction vector and the mean curvature. A topological fusion operation is performed on the first topological feature tensor and the second topological feature vector to generate a node vulnerability topological fingerprint.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention solves the problem that traditional deep learning methods in the prior art require a large number of fault samples for training by recognizing abnormal patterns based on abrupt changes in phase interaction coupling strength and combining the few-sample learning capability of meta-learning algorithms. This enables effective identification of weak links even under the realistic condition of scarce fault samples in power systems, and improves the engineering practicality and application value of the method.

[0017] This invention overcomes the problem that existing technologies, which use fixed parameters and thresholds, cannot adapt to changes in system operating conditions by constructing a system state manifold space and establishing a closed-loop feedback adjustment mechanism for manifold distortion characteristics. When a local distortion of the manifold is detected, it can automatically adjust key parameters such as the phase jump discrimination threshold and the time window length, thereby achieving adaptive optimization of the recognition algorithm and improving the recognition accuracy and robustness in complex and variable power grid environments.

[0018] This invention generates a node vulnerability topological fingerprint by tensor fusion of the adjusted coupling strength mutation feature and the geometric topological feature of the manifold distortion, and identifies node clusters based on the isomorphism matching of the topological fingerprint. This solves the problems of the existing technology's single feature representation and inability to integrate local anomalies and global state information, and realizes multi-scale information integration from micro-node features to macro-system state, which significantly improves the accuracy and comprehensiveness of weak link identification. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0020] Figure 1 This is a flowchart of a novel deep learning-based method for identifying weak links in a power system according to the present invention. Detailed Implementation

[0021] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make the description of this application more complete and comprehensive, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The drawings are merely illustrative illustrations of this application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0022] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more exemplary embodiments. Numerous specific details are provided in the following description to give a full understanding of the exemplary embodiments disclosed in this application. However, those skilled in the art will recognize that the technical solutions disclosed in this application can be practiced with one or more specific details omitted, or other methods, components, steps, etc., can be employed. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the disclosure of this application.

[0023] Example 1 like Figure 1 As shown, this embodiment discloses a novel method for identifying weak links in a power system based on deep learning, including: S101: Collects the active and reactive power of each node in the power system; identifies abnormal energy transmission patterns of nodes by abrupt changes in the phase interaction coupling strength between active and reactive power. It should be noted that active power Reactive power represents the actual electrical power transmitted by node i at time t, reflecting the actual consumption or generation of electrical energy; This represents the reactive power exchange at node i at time t, reflecting the periodic exchange of electromagnetic field energy. Data acquisition is achieved through a synchronous phasor measurement unit or a digital power quality analyzer to ensure the time synchronization of measurement data at each node.

[0024] Understandably, the physical mechanism for identifying abrupt changes in phase interaction coupling strength is based on the discontinuous changes in the phase relationship between nodes during power system faults or disturbances. The active and reactive power are represented as a complex power vector in complex form:

[0025] Where j is the imaginary unit; the phase angle of the power complex vector is... The phase interaction coupling strength between nodes i and j is defined as:

[0026] It should be understood that when a power system experiences a disturbance, the phase relationship between nodes will suddenly jump, causing a discontinuous change in the coupling strength. This abrupt change is a key feature for identifying energy transmission anomalies.

[0027] In specific implementation, the method of identifying abnormal node energy transmission modes through abrupt changes in the phase interaction coupling strength between active and reactive power includes: Construct a complex power vector from active power and reactive power, and establish a power complex vector time series. In practical implementation, for nodes in the power system, a power complex vector time series is established over M consecutive sampling times:

[0028] Wherein, the power complex vector at each time step It contains complete power information for that node. For example, the time series length M is determined according to the analysis requirements, and it is necessary to balance the capture of system dynamic characteristics with computational complexity.

[0029] A phase jump detection algorithm is used to identify phase jump events in the power complex vector time series, and a set of phase jump events is obtained. It should be noted that the core principle of the phase difference jump detection algorithm is based on the continuous characteristic of phase changes during normal operation; the algorithm first calculates the phase angle difference between adjacent moments:

[0030] Under normal circumstances, the phase angle difference follows a normal distribution with a mean close to zero. However, when anomalies occur, significant jumps in value will appear. To address the phase angle... Periodicity necessitates phase expansion processing: when hour, ; when hour,

[0031] Specifically, the step of identifying phase jump events in the power complex vector time series using the phase difference jump detection algorithm includes: Calculate the phase angle difference between adjacent time steps of the power complex vector time series, and generate a phase angle difference sequence; It should be understood that the calculation of the phase angle needs to consider the quadrant of the arctangent function; for complex power vectors... Phase angle The calculation requires determining the correct quadrant based on the signs of active and reactive power, because the standard arctan function only covers... scope.

[0032] In practical implementation, the generated phase angle difference sequence is as follows:

[0033] The sequence has a length of M-1, and each element reflects the phase change characteristics at the corresponding time.

[0034] Establish a statistical distribution model for the phase angle difference sequence and determine the discrimination threshold for phase jump; For example, assume that the phase angle difference sequence during normal operation follows a normal distribution. The distribution parameters are determined by maximum likelihood estimation, where the mean is... Calculated by arithmetic mean Calculated using the sample variance formula.

[0035] It should be noted that the phase jump discrimination threshold is adopted. The criteria are defined as follows:

[0036] This threshold setting is based on the statistical properties of the normal distribution and can effectively identify phase jump events that deviate significantly from the normal range.

[0037] The time corresponding to the phase angle difference exceeding the discrimination threshold is marked as a candidate point for phase change; It is understandable that by examining the phase angle difference sequence one by one, the conditions will be met. The moment These are marked as candidate points for phase abrupt change. The set of candidate points is defined as follows: .

[0038] Verify that the phase continuity before and after the phase change candidate point is indeed interrupted, confirm it as a real phase change event, and output the set of phase change events; It should be understood that phase continuity verification aims to eliminate spurious abrupt changes caused by measurement noise or transient disturbances. For each candidate point... Calculate the phase change rate within the time windows before and after:

[0039] in The time window length; the condition for determining a continuity interruption is that the difference in the rate of change before and after exceeds a preset threshold.

[0040] The final set of phase transition events consists of the validated candidate points:

[0041] Calculate the magnitude of the coupling strength change for each abrupt event in the set of phase abrupt events to form a coupling strength change spectrum; In practice, for each confirmed phase transition moment Calculate the difference in phase interaction coupling strength before and after this moment. The magnitude of the coupling strength change is defined as:

[0042] in and These are the next and previous sampling points at the moment of abrupt change, respectively. Node j traverses all adjacent nodes electrically connected to node i and selects the coupling strength with the largest change as the feature value of node i.

[0043] For example, the spectrum of coupling strength changes consists of the magnitudes of changes in all abrupt events:

[0044] This variation spectrum reflects the distribution characteristics of the perturbation intensity of the system at each abrupt change moment.

[0045] Extract the time window and amplitude features corresponding to abnormal peaks from the coupling strength variation spectrum; It should be noted that abnormal peak extraction is achieved through a local maximum detection algorithm. The algorithm needs to simultaneously satisfy both amplitude and neighborhood conditions: the amplitude condition requires the peak value to be greater than a preset threshold, and the neighborhood condition requires the peak value to be the maximum value within its local neighborhood.

[0046] Understandably, for each detected abnormal peak, its time window features and amplitude features need to be extracted; the time window length... The duration of the anomalous event is determined by the full width at half maximum (FWHM) of the peak value. Amplitude characteristics include peak height. Peak area and peak steepness It comprehensively describes the intensity characteristics of abnormal events.

[0047] Construct node energy transmission anomaly patterns based on the time window and amplitude characteristics of abnormal peak values; It should be understood that the node energy transmission anomaly pattern constructs a multi-dimensional feature vector by fusing time window features and amplitude features; for node i, its anomaly pattern is represented as:

[0048] In practical implementation, amplitude characteristics include peak height. Peak area and peak steepness Among them, peak height The maximum amplitude value of the peak value; where, the peak steepness is defined as... , The peak baseline height is determined by averaging the local minimums before and after the peak. To achieve The time point at which the peak begins For the first time, the magnitude exceeded The point in time.

[0049] For example, this multidimensional feature representation can comprehensively characterize the time and frequency domain features of node energy transmission anomalies, providing rich input information for subsequent deep learning models and ensuring that the models can accurately identify and classify different types of abnormal events.

[0050] S102: Based on a small number of historical disturbance samples, a deep learning model for identifying weak links is generated by training a meta-learning algorithm; the abnormal energy transmission pattern of nodes is input into the deep learning model for identifying weak links to obtain the node response path activation sequence. It should be noted that a small number of historical disturbance samples refers to a limited number of fault or disturbance events recorded during the operation of the power system, typically containing 10 to 50 disturbance samples of different types. Each sample includes the time, location, type, and system response data of the disturbance. Because power system fault events are relatively rare and acquisition is costly, meta-learning algorithms can learn general recognition capabilities from a limited number of samples.

[0051] Understandably, the core idea of ​​meta-learning algorithms is "learning to learn," that is, enabling deep learning models to quickly adapt to new tasks through training on multiple related tasks. Unlike traditional deep learning, which requires a large number of samples, meta-learning algorithms are specifically designed to solve the problem of learning with few samples.

[0052] In the specific implementation process, the step of training a deep learning model for identifying weak links using a meta-learning algorithm based on a small number of historical disturbance samples includes: A small number of historical perturbation samples are divided into a support set and a query set to construct a meta-learning training task. It should be understood that the division of the support set and the query set is a fundamental setting of the meta-learning algorithm.

[0053] support set Contains a small number of samples for each type of perturbation, used for rapid model adaptation; query set The remaining samples, including those for each type of perturbation, are used to evaluate the model's generalization performance.

[0054] In practical implementation, in the K-shot N-way meta-learning setup, each training task contains N randomly selected perturbation classes. Each class contains K samples in the support set and the remaining MK samples in the query set, where M is the total number of samples for that class; the support set... The query set contains N×K samples. It contains N×(MK) samples; the meta-learning training task is defined as:

[0055] in, This represents the i-th training task; For example, in the K-shot N-way setup, each task contains N categories, and each category has K samples in the support set.

[0056] Design a two-layer optimized meta-learning network, which includes a fast-adaptive inner deep network and a cross-task generalizing outer deep network; It should be noted that the two-layer optimization structure is the core architecture of the meta-learning algorithm. The parameters of the inner deep network are denoted as... The outer deep network is responsible for quickly learning feature representations for a specific task on the support set; the meta-parameters of the outer deep network are denoted as... They are responsible for learning general knowledge across tasks.

[0057] Understandably, the inner network employs a convolutional neural network structure, comprising three convolutional layers (3×3 kernel size, 32, 64, and 128 feature maps respectively), two max-pooling layers (2×2 pooling windows), and two fully connected layers (256 neurons, [number of categories]). The outer meta-network updates its meta-parameters using the gradient descent algorithm. .

[0058] Train the parameters of the inner deep network on the support set and optimize the meta-parameters of the outer deep network on the query set; It should be understood that the core of bi-level optimization lies in the alternating update mechanism of the inner and outer layer parameters. The formula for updating the inner layer parameters is:

[0059] in The inner learning rate, Let be the loss function. The formula for updating the outer layer meta-parameters is:

[0060] in Let T be the outer learning rate, and T be the total number of training tasks.

[0061] In practice, the loss function typically employs either cross-entropy loss or mean squared error loss, depending on the specific recognition task. Gradient calculation is performed using the backpropagation algorithm, which requires calculating the partial derivatives of the loss function with respect to the parameters.

[0062] Through multiple rounds of meta-training iterations, the two-layer deep network is equipped with the ability to quickly learn and identify weak points from a small number of new samples; For example, the meta-training process consists of multiple training epochs, in which different combinations of tasks are randomly sampled for training. During training, the average loss on the query set needs to be monitored, and training stops when the loss converges or the preset maximum number of training epochs is reached.

[0063] It should be noted that the rapid learning ability is achieved through training on a large number of tasks. During training, the network learns how to extract effective features from a small number of samples and how to quickly adjust parameters to adapt to new types of perturbations.

[0064] Save the parameters and structure configuration of the two-layer deep network after training convergence to generate a deep learning model for weak link identification. Understandably, a trained deep learning model for identifying weak points includes a complete network structure definition, pre-trained parameter weights, and relevant hyperparameter configurations. The model is typically saved in a standard deep learning framework format for easy loading and use later.

[0065] In the specific implementation process, the step of inputting the abnormal node energy transmission pattern into the weak link identification deep learning model to obtain the node response path activation sequence includes: Convert the node energy transfer anomaly pattern into the input tensor format of the weak link identification deep learning model; It should be understood that there are abnormal modes of node energy transfer. It needs to be converted to a tensor format that can be processed by deep learning models; the conversion process includes dimension reshaping, data type conversion, and batch dimension addition.

[0066] In practical implementation, the feature vector of the abnormal pattern It was reshaped into a four-dimensional tensor, with the tensor dimension being... ,in This is the batch size, typically set to 1 for processing on a single node.

[0067] Extract multi-level feature representations of abnormal patterns through convolutional and pooling layers of deep learning models; It should be noted that convolutional layers perform convolution operations on the input features using a sliding window to extract local feature patterns. Pooling layers reduce the spatial dimensionality of the feature map through downsampling operations while retaining the most important feature information.

[0068] Specifically, the extraction of multi-level feature representations of abnormal patterns through convolutional and pooling layers of a deep learning model includes: The input tensor is passed through the first convolutional layer to perform feature convolution operations, generating a primary feature map. Understandably, the first convolutional layer's role is to extract basic features from the original input. The convolution operation formula is:

[0069] in For input signal, The kernel is a convolution kernel; the generated primary feature map contains local feature representations of the input data.

[0070] Perform max pooling on the primary feature map to reduce the feature dimension while preserving salient features; For example, max pooling selects the maximum value as the output within each pooling window, achieving dimensionality reduction while preserving the most salient feature information. Pooling operations help enhance the model's translation invariance and reduce the risk of overfitting.

[0071] The pooling results are then passed through a second convolutional layer for deep feature extraction to generate a high-level feature map. It should be understood that the second convolutional layer further extracts more abstract, high-level features based on the primary features. These high-level feature maps are able to capture more complex patterns and relationships within the input data.

[0072] The high-level feature map is flattened into a one-dimensional feature vector, and the output is a multi-level feature representation of the anomaly pattern. In practice, the high-level feature map is converted into a one-dimensional vector through a flattening operation. This vector contains multi-level feature information from low to high levels, providing rich feature input for subsequent sequence modeling.

[0073] Multi-level feature representations are input into the recurrent neural network layers of the model to predict possible response propagation paths; It should be noted that this model adopts a CNN-RNN hybrid architecture, with the front-end CNN module extracting one-dimensional feature vectors. Forming a sequence in the time dimension This time series serves as the input to the RNN module. The recurrent neural network layer (usually LSTM or GRU) is used to model the temporal dependencies and propagation paths between nodes. By learning historical perturbation propagation patterns, the network can predict the possible propagation paths of current anomalous patterns.

[0074] Calculate the activation probability and activation intensity of nodes on each predicted path; Understandably, the activation probability is calculated using the softmax function to ensure that the sum of the probabilities of all paths is 1. The activation strength value reflects the degree of impact of the perturbation on each node and is usually determined based on the path weights and the importance of the node in the network.

[0075] The predicted path and activation information are combined to form a node response path activation sequence; It should be understood that the node response path activation sequence is the final output, and its data structure is as follows:

[0076] in, For path identifiers, The sequence of nodes on the path, This is the activation probability sequence for each node. This is the sequence of activation intensity values. This sequence provides important input data for the subsequent construction of the system state manifold space.

[0077] S103: Construct the system state manifold space based on the temporal evolution characteristics of the node response path activation sequence; when the manifold space is locally distorted, use the distortion characteristics to feedback and adjust the abrupt identification of the phase interaction coupling strength in step S101. It should be noted that the system state manifold space is a low-dimensional geometric structure used to represent the operating state of a power system at different times. The construction of the manifold space is based on the time-domain evolution characteristics of the node response path activation sequence, which can map high-dimensional system state data into a low-dimensional space that is easy to analyze and visualize.

[0078] Understandably, when a power system experiences anomalies or disturbances, its representation in the manifold space deviates from its normal operating trajectory, resulting in local distortions. These distortions contain crucial information about the system anomalies and can be used to adjust and optimize disturbance identification algorithms.

[0079] In the specific implementation process, the construction of the system state manifold space based on the temporal evolution characteristics of the node response path activation sequence includes: Collect node response path activation sequences at multiple consecutive time points to form an activation sequence time-domain dataset; It should be understood that constructing a time-domain dataset requires collecting node response path activation sequences within a continuous time window. Let the time window length be T, and the sampling interval be... Then you can obtain The activation sequence data at each time point.

[0080] In specific implementation, the activation sequence time-domain dataset is defined as:

[0081] in, Indicates time The node response path activation sequence; each activation sequence contains multi-dimensional data such as path information, node sequence, activation probability, and activation intensity.

[0082] The activation sequence time-domain dataset is mapped to a low-dimensional manifold space using a manifold embedding algorithm to obtain a low-dimensional manifold space coordinate representation. Specifically, the step of mapping the activation sequence time-domain dataset to a low-dimensional manifold space using a manifold embedding algorithm includes: Construct a proximity graph among data points in the time-domain dataset of the activation sequence; For example, manifold embedding algorithms employ nonlinear dimensionality reduction methods such as Locally Linear Embedding (LLE) or Isomap. The core idea of ​​these algorithms is to map high-dimensional data to a low-dimensional space while preserving the local geometric structure of the data.

[0083] It should be noted that the manifold embedding process first constructs a proximity graph between data points, then calculates the Laplacian matrix of the graph, and finally obtains the low-dimensional embedding coordinates through eigenvalue decomposition.

[0084] Understandably, the construction of a proximity graph is based on a similarity metric between data points. For any two time-step activation sequences... and Calculate the Euclidean distance or cosine similarity between the data points. When the similarity exceeds a preset threshold, establish a connection edge between the corresponding data points.

[0085] Calculate the Laplacian matrix of the proximity graph, perform eigenvalue decomposition, and obtain the eigenvalue vector and eigenvector matrix; It should be understood that the graph Laplacian matrix L is defined as the difference between the degree matrix D and the adjacency matrix A: ; Eigenvalue decomposition produces eigenvalues and the corresponding feature vector .

[0086] Select the eigenvectors corresponding to the smallest non-zero eigenvalues ​​as the embedding basis vectors; In practice, the eigenvectors corresponding to the first d smallest non-zero eigenvalues ​​are usually selected to form the embedding basis, where d is the dimension of the target low-dimensional space, usually taken as 2 or 3 to facilitate visualization and analysis.

[0087] The original data points are projected onto the subspace spanned by the embedded basis vectors to obtain a low-dimensional manifold space coordinate representation; For example, coordinates in a low-dimensional manifold space are represented as:

[0088] Where X is the original high-dimensional data matrix and Y is the low-dimensional embedded coordinate matrix.

[0089] Identify the distribution patterns and clustering structures of data points in a low-dimensional manifold space; It should be noted that the distribution pattern of data points reflects the characteristics of the power system under different operating conditions. Data points under normal operating conditions usually cluster in a certain region of the manifold space, forming a dense cluster structure; data points under abnormal conditions deviate from the normal clusters, forming outliers or new clusters.

[0090] Understandably, cluster structure identification can employ density-based clustering algorithms (such as DBSCAN) or hierarchical clustering algorithms. These algorithms can automatically identify cluster regions of different densities without requiring a pre-specified number of clusters.

[0091] Extract the geometric boundaries and internal connectivity of the cluster structure to construct the system state manifold space; It should be understood that geometric boundaries define the dividing lines between different operating state regions, and internal connectivity describes the correlation between data points within the same cluster. By extracting these geometric features, a complete representation of the system state manifold space can be constructed.

[0092] In practical implementation, the system state manifold space It includes geometric descriptors such as cluster centers, boundary information, and connection weights, providing a foundation for subsequent distortion detection.

[0093] It should be noted that local distortion of the manifold space refers to a significant change in the geometry of clustered regions under normal operating conditions, including phenomena such as cluster center shift, boundary expansion or contraction, and internal structural deformation. These changes usually indicate anomalies in the operating state of the power system.

[0094] Understandably, distortion detection is achieved by comparing the current manifold structure with the historical baseline structure. When a significant geometric change is detected, the system extracts distortion features and feeds them back to step S101 to adjust the identification parameters for abrupt changes in phase interaction coupling strength.

[0095] In practical implementation, the mathematical implementation of the feedback adjustment mechanism is as follows: First, calculate the manifold distortion intensity. ,in This is the current manifold structure matrix. As the baseline manifold structure matrix, The Frobenius norm is used; then the recognition parameters are adjusted based on the torsion intensity: the adjusted phase jump discrimination threshold. Adjusted time window length ,in, and The adjustment coefficients are determined through regression analysis of historical disturbance data; when Time-triggered parameter adjustment, among which The distortion threshold.

[0096] S104: Generate node vulnerability topological fingerprints based on the geometric topological features of the adjusted coupling strength abrupt change and manifold distortion; identify node clusters with the same vulnerable topological structure as weak links in the system by isomorphic matching of the topological fingerprints; It should be understood that node vulnerability topological fingerprinting is a multi-dimensional feature representation that integrates adjusted coupling strength abrupt changes and the geometric features of manifold distortion. This fingerprint can uniquely identify the vulnerability features of each node, providing a precise mathematical description for weak link identification.

[0097] It should be noted that isomorphism matching refers to identifying clusters of nodes with similar topological characteristics. Topological isomorphism means that the connections and geometric features between nodes remain unchanged under certain transformations, and these nodes often have similar vulnerability patterns.

[0098] In the specific implementation process, generating node vulnerability topological fingerprints based on the adjusted coupling strength abrupt change and manifold distortion geometric topological features includes: Extract the amplitude and direction vectors of the abrupt change in the phase interaction coupling strength after adjustment; In practice, the adjusted phase interaction coupling strength abrupt change data includes information on coupling strength changes optimized by feedback. Amplitude vector The vector represents the magnitude and direction of each mutation event at node i. This indicates the phase direction characteristic of the abrupt change.

[0099] Understandably, the magnitude vector and direction vector together describe the complete information about node energy transmission anomalies, providing basic data for subsequent topological feature calculations.

[0100] Obtain the Gaussian curvature and mean curvature values ​​of the manifold distortion region; For example, Gaussian curvature K and mean curvature H are important parameters for describing the local geometric properties of a manifold. Gaussian curvature reflects the inherent curvature of the manifold at that point, while mean curvature reflects the external curvature characteristics of the manifold.

[0101] It should be noted that curvature calculation is based on the local tangent plane and normal vector of the manifold space. The first and second partial derivatives of the manifold surface are calculated using numerical differentiation methods, thereby obtaining the first and second fundamental forms, and finally calculating the curvature value.

[0102] Calculate the tensor product of the magnitude vector and the Gaussian curvature to obtain the first topological feature tensor; It should be understood that the tensor product operation will convert the magnitude vector... Combined with the Gaussian curvature K, the first topological feature tensor is generated. This operation integrates the perturbation intensity information of the nodes and the geometric curvature information of the manifold.

[0103] The second topological feature vector is obtained by calculating the outer product of the direction vector and the mean curvature. Understandably, the outer product operation will convert the direction vector... Combined with the mean curvature H, a second topological feature vector is generated. This operation captures the correlation between the perturbation direction and the manifold bending direction.

[0104] Perform a topological fusion operation on the first topological feature tensor and the second topological feature vector to generate a node vulnerability topological fingerprint. In practice, topology fusion operations combine two tensor structures of different orders into a unified topological fingerprint representation through specific mathematical transformations. The resulting node vulnerability topological fingerprint... It contains complete vulnerability information of the node.

[0105] For example, by calculating the vulnerability topological fingerprint of all nodes and performing isomorphism matching analysis, node clusters with similar vulnerability structures can be identified. These node clusters are the weak links in the power system and need to be monitored and protected. In the specific implementation process, the topological fusion operation between the first topological feature tensor and the second topological feature vector is as follows: The second topological feature vector is expanded into a tensor form of the same order as the first topological feature tensor. It should be noted that tensor order matching is a prerequisite for topological fusion operations. Second topological feature vector. It needs to be transformed into a tensor with the first topological feature through tensor expansion operation. Same order.

[0106] In practice, vector expansion is implemented using Kronecker product or outer product operations. Let the first topological feature tensor be... If the second topological eigenvector is a second-order tensor, then the second topological eigenvector is expanded into a second-order tensor form by the cross product with the unit vector.

[0107] Perform the Hadamard product operation on the two tensors to obtain the preliminary fused tensor; It is understandable that the Hadamard product is a tensor operation involving element-wise multiplication. Let the expanded second topological feature tensor be... The initial fusion tensor is then:

[0108] Where ⊙ represents the Hadamard product operation. This operation preserves the local feature information of the two tensors; Singular value decomposition is performed on the initial fusion tensor to extract the singular vectors corresponding to the main singular values. It should be understood that Singular Value Decomposition (SVD) decomposes the fused tensor into the product of three matrices:

[0109] Where U and V are orthogonal matrices, and Σ is a diagonal matrix. By retaining the largest singular values ​​and their corresponding singular vectors, the main eigenvalues ​​of the tensor can be obtained.

[0110] The main singular vectors are combined to form a topological fingerprint of node vulnerability; For example, in a specific implementation, singular values ​​greater than a threshold are selected. Singular vectors are taken as principal components, and it is assumed that they satisfy the condition If the number of singular values ​​is k, then the node vulnerability topological fingerprint is:

[0111] in, Let j be the j-th principal left singular vector; , The maximum singular value is the topological fingerprint; this topological fingerprint contains the most important characteristic information about node vulnerability.

[0112] It should be noted that isomorphism matching is based on a similarity measure between topological fingerprints. The more similar the topological fingerprints of two nodes, the more similar their vulnerability characteristics are, and the more likely they are to become weak points in the system.

[0113] In practical implementation, isomorphic matching of topological fingerprints identifies clusters of nodes with the same vulnerable topological structures as weak links in the system, including: Calculate the isomorphic similarity between the vulnerability topological fingerprints of each node. Understandably, topological isomorphism is achieved through a comprehensive evaluation of structural and geometric similarity; the topological fingerprint of nodes i and j and Isomorphism similarity is calculated as follows:

[0114] in, For cosine similarity, For structural similarity, and α is the weighting coefficient and α+β=1, σ is the bandwidth parameter.

[0115] Establish a threshold discrimination criterion for topological fingerprint similarity; It should be understood that the threshold discrimination criteria need to balance recognition accuracy and coverage. A threshold that is too high may cause potential weak points to be missed, while a threshold that is too low will result in too many false recognitions.

[0116] In practice, the threshold can be determined through statistical analysis methods, such as setting it as the mean of all similarity values ​​plus one standard deviation, or determining the optimal threshold point through ROC curve analysis.

[0117] Based on the threshold discrimination criterion, nodes with similarity higher than the threshold are marked as isomorphic node pairs. For example, the set of isomorphic node pairs is defined as:

[0118] in This is the preset similarity threshold.

[0119] Finally, cluster analysis was performed on homogeneous node pairs to form node clusters with the same vulnerability topology as weak links in the system. It should be noted that cluster analysis organizes nodes with high similarity into connected cluster structures; graph clustering algorithms can be used, with nodes as vertices of the graph and connections between isomorphic node pairs as edges, and community detection algorithms can be used to identify tightly connected node clusters.

[0120] Understandably, each identified cluster of nodes represents a region in the power system with similar vulnerability characteristics. These regions may exhibit similar response patterns when faced with disturbances, and therefore need to be monitored and protected as a whole.

[0121] In practice, the final output of system weaknesses includes information such as the identification of node clusters, a list of nodes within the cluster, the vulnerability rating of the cluster, and recommended protective measures, providing decision support for the operation and maintenance of the power system.

[0122] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters, weights, and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0123] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0124] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A novel method for identifying weak links in power systems based on deep learning, characterized in that, include: S101: Collect the active and reactive power of each node in the power system; Abnormal energy transfer patterns at nodes can be identified by abrupt changes in the phase interaction coupling strength between active and reactive power. S102: Based on a small number of historical disturbance samples, a deep learning model for identifying weak links is generated by training a meta-learning algorithm. Input the abnormal energy transmission pattern of nodes into the deep learning model for weak link identification to obtain the node response path activation sequence. S103: Construct the system state manifold space based on the temporal evolution characteristics of the node response path activation sequence; when the manifold space is locally distorted, use the distortion characteristics to feedback and adjust the abrupt identification of the phase interaction coupling strength in step S101. S104: Generate node vulnerability topological fingerprints based on the geometric topological features of the adjusted coupling strength abrupt change and manifold distortion; identify node clusters with the same vulnerable topological structure as weak links in the system by isomorphic matching of the topological fingerprints.

2. The novel deep learning-based method for identifying weak links in power systems according to claim 1, characterized in that, The method of identifying abnormal energy transfer patterns at nodes by abrupt changes in the phase interaction coupling strength between active and reactive power includes: Construct a complex power vector from active power and reactive power, and establish a power complex vector time series. A phase jump detection algorithm is used to identify phase jump events in the power complex vector time series, and a set of phase jump events is obtained. Calculate the magnitude of the coupling strength change for each abrupt event in the set of phase abrupt events to form a coupling strength change spectrum; Extract the time window and amplitude features corresponding to abnormal peaks from the coupling strength variation spectrum; Anomaly patterns in node energy transmission are constructed based on the time window and amplitude characteristics of abnormal peak values.

3. The novel deep learning-based method for identifying weak links in power systems according to claim 2, characterized in that, The method of identifying phase jump events in the power complex vector time series using the phase difference jump detection algorithm includes: Calculate the phase angle difference between adjacent time steps of the power complex vector time series, and generate a phase angle difference sequence; Establish a statistical distribution model for the phase angle difference sequence and determine the discrimination threshold for phase jump; The time corresponding to the phase angle difference exceeding the discrimination threshold is marked as a candidate point for phase change; Verify that the phase continuity before and after the phase change candidate point is indeed interrupted, confirm it as a real phase change event, and output the set of phase change events.

4. The novel deep learning-based method for identifying weak links in power systems according to claim 3, characterized in that, The process of training a deep learning model for identifying weak points using a meta-learning algorithm based on a small number of historical disturbance samples includes: A small number of historical perturbation samples are divided into a support set and a query set to construct a meta-learning training task. Design a two-layer optimized meta-learning network, which includes a fast-adaptive inner deep network and a cross-task generalizing outer deep network; Train the parameters of the inner deep network on the support set and optimize the meta-parameters of the outer deep network on the query set; Through multiple rounds of meta-training iterations, the two-layer deep network is equipped with the ability to quickly learn and identify weak points from a small number of new samples; Save the parameters and structure configuration of the two-layer deep network after training convergence to generate a deep learning model for weak link identification.

5. The novel deep learning-based method for identifying weak links in power systems according to claim 4, characterized in that, The step of inputting the abnormal node energy transmission pattern into the weak link identification deep learning model to obtain the node response path activation sequence includes: Convert the node energy transfer anomaly pattern into the input tensor format of the weak link identification deep learning model; Extract multi-level feature representations of abnormal patterns through convolutional and pooling layers of deep learning models; Multi-level feature representations are input into the recurrent neural network layers of the model to predict possible response propagation paths; Calculate the activation probability and activation intensity of nodes on each predicted path; The predicted path and activation information are combined to form the node response path activation sequence.

6. The novel deep learning-based method for identifying weak links in power systems according to claim 5, characterized in that, The extraction of multi-level feature representations of abnormal patterns through convolutional and pooling layers of a deep learning model includes: The input tensor is passed through the first convolutional layer to perform feature convolution operations, generating a primary feature map. Perform max pooling on the primary feature map to reduce the feature dimension while preserving salient features; The pooling results are then passed through a second convolutional layer for deep feature extraction to generate a high-level feature map. The high-level feature map is flattened into a one-dimensional feature vector, and the output is a multi-level feature representation of the anomaly pattern.

7. The novel deep learning-based method for identifying weak links in power systems according to claim 6, characterized in that, The construction of the system state manifold space based on the temporal evolution characteristics of node response path activation sequences includes: Collect node response path activation sequences at multiple consecutive time points to form an activation sequence time-domain dataset; The activation sequence time-domain dataset is mapped to a low-dimensional manifold space using a manifold embedding algorithm to obtain a low-dimensional manifold space coordinate representation. Identify the distribution patterns and clustering structures of data points in a low-dimensional manifold space; Extract the geometric boundaries and internal connectivity of the cluster structure to construct the system state manifold space.

8. The novel deep learning-based method for identifying weak links in power systems according to claim 7, characterized in that, The process of mapping the activation sequence time-domain dataset to a low-dimensional manifold space using a manifold embedding algorithm includes: Construct a proximity graph among data points in the time-domain dataset of the activation sequence; Calculate the Laplacian matrix of the proximity graph, perform eigenvalue decomposition, and obtain the eigenvalue vector and eigenvector matrix; Select the eigenvectors corresponding to the smallest non-zero eigenvalues ​​as the embedding basis vectors; The original data points are projected onto the subspace spanned by the embedded basis vectors to obtain a low-dimensional manifold space coordinate representation.

9. The novel deep learning-based method for identifying weak links in power systems according to claim 8, characterized in that, The generation of node vulnerability topological fingerprints based on the adjusted coupling strength abrupt change and manifold distortion geometric topological features includes: Extract the amplitude and direction vectors of the abrupt change in the phase interaction coupling strength after adjustment; Obtain the Gaussian curvature and mean curvature values ​​of the manifold distortion region; Calculate the tensor product of the magnitude vector and the Gaussian curvature to obtain the first topological feature tensor; The second topological feature vector is obtained by calculating the outer product of the direction vector and the mean curvature. A topological fusion operation is performed on the first topological feature tensor and the second topological feature vector to generate a node vulnerability topological fingerprint.

Images